Improved Prostate Tumor Segmentation Using Self-Attention Mechanism and SAM Model

Accurate segmentation of prostate tumors from medical images is crucial for diagnosis and treatment planning in prostate cancer patients. In this research, we propose an improved approach for prostate tumor segmentation leveraging the Self-Attention Mechanism (SAM) and fine-tuning techniques. Our method involves several key steps: first, we preprocess the medical images and masks, converting them into suitable formats for processing. Next, we employ a deep learning model, specifically the SAM - ViT model pretrained on large-scale datasets, to perform segmentation. We fine-tune this model on our dataset to adapt it to the specific characteristics of prostate tumor segmentation. To enhance the accuracy of segmentation, we introduce a novel technique for adjusting the threshold in the post-processing step, which significantly impacts the quality of the segmentation masks. We optimize this threshold dynamically during inference to achieve the best Intersection over Union (IoU) scores. Furthermore, we evaluate our method on both training and test datasets, demonstrating its effectiveness in accurately segmenting prostate tumors. Our results indicate notable improvements in segmentation quality compared to existing methods, as evidenced by higher IoU scores. Overall, our proposed approach offers a promising solution for enhancing prostate tumor segmentation accuracy, which could potentially improve clinical decision-making and patient outcomes in prostate cancer management.